Ruminant Monitoring System

Abstract

A system and method for automatically monitoring ruminants is provided. The system includes a 3D camera system that obtains images from a region of interest, particularly a lumbar fossa. An image processor determines the surface curvature in the region of interest as a function of time. Based on the frequency with which the function reaches a local maximum, an indication of the health of the animal is generated.

Description

[0001] The present invention relates to a system for automatically monitoring ruminants, in particular dairy animals such as cows, the system comprising: a 3D camera system for obtaining multiple 3D images of at least one area of ​​interest of the animal at consecutive time points during at least a predetermined time period; a control device connected to the 3D camera system, the control device being provided with an image processing device for processing the multiple 3D images obtained; and an output device.

[0002] Such animal monitoring systems are known. For example, document WO 12 / 138290 A1 discloses a system for determining the intestinal filling level of the rumen of dairy animals. A 3D camera system determines the depth or volume of the rumen triangle. A body condition scoring system based on a 3D camera is also known.

[0003] However, the gut filling level or body condition score has limited importance for the health status of ruminants.

[0004] Therefore, there is a need for practical, simple, and reliable systems and methods for determining more and / or better health-indicating information about ruminants based on 3D images of the animals.

[0005] The present invention at least partially achieves this goal and in one aspect provides a system for automatically monitoring ruminants, in particular dairy animals such as cows, the system comprising: a 3D camera system for obtaining multiple 3D images of at least one region of interest of the animal at consecutive time points during at least a predetermined time period; a control device connected to the 3D camera system, the control device being provided with an image processing device for processing the obtained multiple 3D images; and an output device, the control device being arranged to determine a health indication based on the processed image and output the health indication to the output device, wherein the image processing device is arranged to: determine a region of interest in the multiple 3D images, and one of the following: calculate a curvature value of the region of interest for each 3D image and determine the calculated curvature as a function of time, or measure a relative position of at least a predetermined point in the region of interest relative to the animal for each 3D image and determine the relative position as a function of time, wherein the control device is arranged to determine a time point at which a local extreme value of the time function occurs and the health indication (this is performed by analyzing the time point relative to a predetermined standard).

[0006] The following insight is used herein: Knowing the motility of the rumen-reticulum can provide valuable health information. The rumen-reticulum performs digestive functions by periodic contractions and relaxations. In this context, the initial contractions start from the reticulum and pass through the rumen. These contractions mix and circulate the chyme. These contractions can be observed by a 3D camera. It is noted here that determining only a single value of the curvature (which value can be derived, for example, from prior art systems that determine the depth or volume of the rumen area or body condition scores from a single image) does not yield any information about the contractions.

[0007] Note that rumen motility is often determined by a veterinarian through auscultation (listening to body sounds) and tremors (diagnosing by feeling the body). This requires the presence of a veterinarian, which is impractical for continuous monitoring. Therefore, in practice, rumen motility is only used when some health problem is suspected. Therefore, the possibility of determining health problems or their precursors at a very early stage is eliminated. In addition, actual tests have shown that it is more difficult for humans to observe the contractions of a filled rumen than can be achieved using the system according to the present invention, which therefore provides advantages in terms of reliability. Since it is a manual assessment, the veterinarian's determination of motility may also be subject to errors and subjectivity. And this may also cause stress or other disturbances to the animal or herd. This itself may be the cause of deviations from the true rumen-reticulum motility, so that the veterinarian may make an erroneous assessment due to stress caused by his own assessment, etc. For these reasons, it is desirable to have a non-contact system for determining health indicators as provided by the present invention.

[0008] In the method, local extrema of a function can be determined as, in particular, a calculated curvature value or a local maximum in time of a relative position. The position maximum here is considered to be the position closest to the 3D camera. "Local" is considered to be "relative to a variable" in a mathematical sense, not necessarily related to a spatial position.

[0009] The images are acquired at consecutive time points. These time points need not be distributed at regular time intervals, but they must be known in order to allow temporal analysis of the images. Furthermore, the images should also be acquired during at least a predetermined time period, again in order to allow a reasonable temporal analysis. This time range depends on the expected recurrence interval, which may depend on the type of animal, etc. In practice, it is preferred if the images are acquired during at least 2 minutes, and more preferably during at least 5 minutes. At least for dairy cows, this allows a sufficient number of contractions to be observed to reliably determine a health indicator value.

[0010] The 3D camera may be fixed or movable, as long as the camera is arranged to observe the region of interest. Furthermore, taking into account that the animal may move during image acquisition, the 3D camera may be provided with a lens having such a field of view that the region of interest will be included in the image for all possible positions of the animal relative to the 3D camera. In particular, the 3D camera may be arranged to be directed to the region of interest by means of a provided mover control connected to a control device. The control device may then be able to actively direct the 3D camera to the region of interest, for example based on image processing. A smaller field of view may then be used, which in turn allows for greater accuracy.

[0011] In practice, determining the region of interest in a plurality of 3D images may advantageously benefit from some form of tracking, particularly when the animal is moving. For example, the image processing device is arranged to identify a particular point in the image and use this point to track and locate consecutive images for comparison. Based on such tracking, the camera's mover control means may be arranged to position the 3D camera so that the identified point is always at the same position in the image.

[0012] Further special and inventive embodiments are described in the dependent claims and in the present following part of the description.

[0013] In an embodiment, the 3D camera system comprises a time-of-flight camera or a structured light camera. This type of 3D camera provides a relatively high frame rate of up to 30 Hz or more. This ensures that the contraction or other movement of the region of interest between images is not missed, and the movement of the animal can be easily taken into account in the 3D image obtained. However, it is not necessarily necessary to have such a high frame rate, which would result in a relatively large number of calculations. Other 3D camera types such as stereo cameras or laser scanners are also possible. Examples of preferred 3D cameras are the "SwissRanger" 3D camera or equivalent, and a "Kinect" type camera device. In all of these, the camera's ability to provide 3D information is used in the present invention to provide curvature data of the region of interest of the animal.

[0014] In an embodiment, the region of interest is the left paralumbar fossa, also known as the rumen triangle. This is a more or less triangular area on the left side of the ruminant that provides a relatively clear view of the rumen reticulum. It can be easily identified automatically, for example by template matching using the triangle and matching areas of relatively high local curvature. A similar method is described at the following URL: https: / / nl.mathworks.com / help / images / ref / normxcorr2.html, and reference is made to this source for mathematical details. In particular, the region of interest is a partial area within the left paralumbar fossa, i.e. not the entire rumen triangle but only a smaller or larger portion thereof. It has been found that useful information can also be gathered by looking at a portion of the paralumbar fossa, which limits the number of determinations and calculations.

[0015] In an embodiment, the curvature value comprises or is the average value of the curvature values ​​of the region of interest, in particular the average value averaged over a plurality of points across the region of interest. The curvature value may be determined according to any desired method, as long as it is an indication of the curvature of the surface. According to this embodiment, the value may be determined by analyzing a complete or partial region of interest. It may also be possible to analyze a plurality of points in the complete or partial region, again as long as a good indication of the curvature is obtained. It should also be noted that the sign of the curvature value cannot be ignored, since a convex region of interest should be given a positive value and a concave region of interest a negative value, or vice versa, as long as the two are different. This may prevent meaningless double peaks in the case where the appearance of the region of interest varies between convex and concave.

[0016] A useful example of calculating the curvature of a surface (region of interest) is by performing a principal component analysis on the covariance matrix of the neighbors of the points on the surface. This gives three eigenvalues ​​of the matrix λ0, λ1, λ2, and the curvature σ of the point can be estimated by σ=λ0 / (λ0+λ1+λ2), where λ0 is the smallest eigenvalue. The surface curvature can then be determined by averaging (or adding, etc.) the surface curvatures of a plurality of points (such as a matrix of points) in the region of interest. It turns out that a grid or grid of surface points around a fixed and easily identifiable point in the region of interest is sufficient, as already explained above. The fixed point can be, for example, the surface point with the highest cross-correlation coefficient used for tracking. However, this does not exclude other methods of estimating the surface curvature, such as by calculating the change of all surface normal vectors in the region of interest or a part / window thereof.

[0017] After determining the surface curvature (this step can be done for each image), the result is the curvature as a function of time. When analyzing this function, local extrema, in particular local maxima, are determined. In this context, it is advantageous to consider a value as a local extrema if the value is the highest in a predetermined time window (such as a few seconds, e.g. 3s) before and after the time of the value. This prevents noise spikes etc. from interfering too much.

[0018] In an embodiment, the control device is arranged to determine the frequency of the local extrema, in particular local maxima, over time and to analyze the time points by comparing the determined frequency with a predetermined frequency criterion. Extrema, in particular maxima, will occur with a more or less regular frequency. By determining this frequency or frequency band, valuable information about the health of the animal can be obtained and a health indication can be given.

[0019] In an embodiment, the control device is arranged to filter the calculated curvature time function by filtering out time variations with frequencies outside a predetermined frequency range. By this measure, irrelevant variations, such as those caused by animal movement / displacement or noise, can be effectively removed. The following insight is used herein: relevant contractions occur within a frequency band specific to the animal. In particular but not limiting embodiments, the control device is arranged to perform a Fourier transform on the calculated curvature function to construct a transformed frequency function, then remove all parts of the transformed frequency function outside the predetermined frequency range to obtain a clean frequency function, and determine the contraction frequency by analyzing the clean frequency function. Fourier transform is a well-known effective method for removing frequency components from a time signal. However, other filtering methods such as bandpass filters can be applied to achieve similar effects. The analysis of the clean frequency function can be completed by determining the frequency value with the strongest signal value or the average value of the frequency signal within the remaining signal in the clean frequency function, or by giving any other analysis indicating the contraction frequency value. In addition, but only optionally, the control device can be arranged to perform an inverse Fourier transform on the clean frequency function to obtain a clean curvature function. The clean curvature function can be displayed for visual inspection.

[0020] In particular, the predetermined frequency range is or includes frequencies from 0.5 to 4 times per minute, inclusive. For cows, this has proven to be a useful frequency band, outside of which there are no signals indicating meaningful contractions. For other species, different frequency bands may be applicable.

[0021] In an embodiment, the frequency criteria include generating a health warning when the determined frequency is below a predetermined frequency threshold. It has been found that for normal animals, the frequency of rumen-reticulum contractions occurs at least at a certain value, and it has also been found that for animals with clinical or even subclinical disease, this frequency decreases. This allows the determination of a health warning, such as an entry on a watch list, a message to the farmer or veterinarian, an audible or visual alarm, etc. All of these serve to draw attention to a particular animal for further health examination because a certain value for the animal was found to be unusual.

[0022] In an embodiment, the frequency threshold is a historical value of the animal, in particular a historical average value of the animal, or the frequency threshold is a literature value of the animal, in particular a literature value depending on the breed, age, number of days of lactation, or one or more of the feed type or feeding regimen of the animal, respectively. This allows fine-tuning the frequency threshold according to the specific situation of the animal in order to prevent unnecessary health warnings. For example, when the animal is mainly fed with roughage, the contraction frequency is usually higher than when fed with a diet rich in concentrate feed, so the threshold can be adjusted accordingly. It should be noted here that the literature value can be a starting point, and optionally processing the historical value by deep learning or the like can produce a threshold with higher reliability.

[0023] In particular, the ruminant is one of a group of ruminants and the frequency threshold depends on the average frequency of the group of ruminants, more particularly, wherein the average is an instantaneous average or a moving average. This ensures that changes applicable to the entire herd can be taken into account without causing an alarm (or health warning). If the herd is allowed to go out to graze on a particular day after staying and feeding indoors for several months, such as in winter, changes in the rumen-reticulum frequency of the entire herd may occur, which is normal. Therefore, it is meaningless to issue an alarm for all animals. However, if this change is significantly different for a certain animal, an alarm may still be issued.

[0024] More generally, it is believed that in virtually all cases, an alarm should be raised for variations relative to the herd average, or at least the possible causes of such deviations should be investigated. In many cases, deviations from the herd average frequency are good indicators of disease, such as metabolic diseases in particular.

[0025] For example, the value of the frequency threshold can be a certain absolute or relative value below or below the herd average frequency, such as 0.5 Hz or 15% above or below the herd average frequency value. However, other values ​​can certainly be selected depending on the type of animal, the size of the herd, the desired accuracy, etc.

[0026] In an embodiment, the frequency threshold depends on the type of activity the animal is performing, such as feeding, milking or resting. The following insight is used herein: the frequency of contractions depends on the type of activity the animal is engaged in over a short period of time. For example, if the animal is eating, the frequency will typically be high, i.e. on the high side within the expected frequency band. Conversely, in a stressful situation or at rest, the frequency will be lower.

[0027] In an embodiment, the system is arranged to determine the rumen filling level based on a plurality of images obtained or at least absolute or relative changes therein. Basically, the same method as in prior art document WO 12 / 138290A1 can be applied, the part of which describing such method is considered to be incorporated. In short, the method and system disclosed in this document determine the average concavity or convexity of the rumen triangle and thereby determine the rumen filling value. Note that in the current embodiment of the invention, it is possible and advantageous to use more than one image in order to determine the average value of the curvature and therefore the average value of the rumen filling. This ensures that the true rumen filling value is correctly determined without being affected or hardly affected by sudden movements, breathing, etc.

[0028] By determining the rumen filling value in this way, the system is able to determine even more accurately when a health warning or alarm can be given. In particular, the control device is arranged to issue a health warning when one of the following occurs:

[0029] - the determined frequency decreases over time and the determined rumen filling value increases over time;

[0030] - after the regime change, the determined frequencies vary more significantly compared to the herd average frequency values;

[0031] - after the change in regime, the determined frequency of the animal does not return to the frequency of that animal before the change in feeding regime; or

[0032] - After the regime change, the determined frequencies of the animals do return to their original values, but their rate of recovery is slower than the herd average by at least a predetermined amount.

[0033] The first of these four possibilities is a sign of suppressed rumen motility, which is a sign of metabolic disease, particularly rumen acidosis, and should be treated.

[0034] The second to fourth possibilities are related to the quantification of metabolic plasticity and the adaptation to institutional changes. Such institutional changes are particularly changes in the feeding system, but may also be changes in (barn) climate, herd density or any farm management practices. In each case, closer monitoring is recommended because the indicated changes are signs of metabolic problems. In this article, it should be noted that to date, this metabolic plasticity has been determined or quantified by laboratory tests of blood values ​​and / or milk. The inventors have found that a similar effective health warning system can be obtained by the present system and method, which is non-invasive and therefore less uncomfortable for the animal.

[0035] The determined frequency and optionally the rumen filling value may also be used as a check on other indicator values. For example, milk production is often used as a health warning criterion. However, a decrease in milk production does not indicate illness in itself, and health warnings based solely on milk production may too often lead to false health warnings or warnings that appear too late and cause permanent deterioration in health or production. By the present invention, a determined rumen-reticulum contraction frequency that is too high or too low (whether in absolute terms or relative to the group average), or a determined frequency that decreases when the rumen filling value increases, according to other combinations mentioned above, etc., may not only cause a health warning in itself, but may actually be the other way around, that is, if milk production decreases and normally causes a health warning, and the determined frequency remains stable, the control unit will overrule the health warning. The control unit is then arranged to give or not give a health warning accordingly. Similarly, if the rumen filling and / or milk fat concentration decreases, but the animal's determined rumen-reticulum contraction frequency remains stable, no health warning is required, and the control unit should also be arranged accordingly. However, a feeding warning may be given, since the latter situation is related to insufficient feed (intake) rather than to an (acute) illness of the animal. In all cases, the system should include any means known in the art for determining additional quantities such as milk production or milk fat concentration.

[0036] In summary, the present invention provides an advantageous method for monitoring the health of ruminants, the method comprising: automatically determining the rumen reticulum contraction frequency and rumen filling by a system according to the present invention, and generating a health warning based on the comparison of the determined rumen reticulum contraction frequency and rumen filling with a predetermined frequency value and a predetermined rumen filling value, respectively. In this context, the corresponding predetermined value can be a time average or a herd average, a literature value, etc. A major advantage of this method is that basing the health warning on these parameters can speed up the detection of serious problems, because when there are problems related to metabolism, rumen contraction and rumen filling change first, followed by changes in body condition score, milk production and milk quality. The description of further system embodiments will now continue.

[0037] In an embodiment, the system comprises at least one station, each station being selected from the group consisting of: a milking station, a feeding station, a drinking station, a treatment station, an isolation station and a selection station, wherein the 3D camera system comprises a 3D camera arranged in at least one of the stations. These stations are very suitable for placing the 3D camera system, because the animals will stay in one position for a relatively long time. In particular, the milking station is very suitable, because it can even be predicted how long the animals will stay there based on the expected milk production and other historical data. On average, the animals will stay at the milking station for at least 5 minutes, which provides a very suitable time window for determining the frequency as already stated above. In addition, such measurements can also be performed using the 3D camera system at other stations. Even if not every instance produces a reliable measurement value (for example because the animal stays in the appropriate position for too short a time), the system of the present invention allows monitoring the animal in many instances, so that a good insight into the health of the animal can be obtained and a warning can be issued as early as possible when problems arise. It should be noted here that at each station, in addition to measuring the rumen-reticulum contraction, some other actions can also be performed, such as milking, feeding, treating or isolating the animal. Advantageously, the quarantine station is coupled to the milking station or feeding station so that in the event of a health warning for a particular animal in the station, the control device can send the animal to the quarantine station to await a health check by the farmer or veterinarian.

[0038] In an embodiment, the control device is arranged to determine the time intervals between local extrema based on the determined time points, and to analyze the time points by comparing the determined time intervals with a predetermined time interval criterion. In the above, the system uses a frequency-based analysis to determine a health indication or warning. However, the analysis can also be performed based on the time intervals between local extrema. In fact, after undergoing a transformation from the time domain to the frequency domain, this analysis will be reduced to the same as the frequency-based analysis. However, in some cases, it may be simpler to analyze only various time intervals, such as by determining the average time interval between local maxima. Such an average time interval should then be within a certain time interval range, which may depend on the animal, the activity, etc. All features relating to a particular embodiment in a frequency-related analysis system also apply to a time interval-based analysis system.

[0039] In a second aspect, the present invention relates to a method for determining a health indication of a ruminant, in particular a dairy cow, using a system according to the first aspect of the invention and comprising the steps of obtaining a plurality of 3D images of at least one region of interest of the animal at successive time points during at least a predetermined time period, processing the plurality of 3D images obtained, determining a region of interest in the plurality of 3D images, one of: calculating a curvature value of the region of interest for each 3D image and determining the calculated curvature as a function of time, or measuring a relative position of at least a predetermined point in the region of interest relative to the animal for each 3D image and determining the relative position as a function of time, and determining the time point at which a local extremum of the time function occurs and a health indication by analyzing the time point relative to a predetermined criterion, and / or determining the amplitude of the function and a health indication by analyzing the determined amplitude relative to a predetermined amplitude criterion. Since this is a method corresponding to the system aspect of the invention, it is sufficient here to state that all the special features and advantages mentioned for the measures related to a special embodiment of the system aspect of the invention also apply to the method aspect.

[0040] It is important to emphasize here that this method is not a diagnostic method, although it can be used in a diagnostic method. After all, this method only helps to determine whether there is a problem with the animal. It cannot make a diagnosis about what disease (if any) is present in the animal that gave the health warning.

[0041] The invention will now be explained with reference to one or more exemplary and non-limiting embodiments and to the accompanying drawings, in which:

[0042] - Figure 1 A system 1 according to the invention is shown very schematically, and

[0043] - Figure 2 Exemplary graphs of raw curvature versus time, and smoothed curvature as a function of time are shown.

[0044] Figure 1 Very schematically, a system 1 according to the invention is shown for determining health indicators for a dairy cow 2 and comprises a 3D camera 3, a control device 4 with an image processing device 5, a milking robot 6 with teat cups 7 and a feeding trough 8 with a sensor 9. A region of interest in the form of the left flank fossa is indicated with reference numeral 10 and a sub-region or window with reference numeral 11.

[0045] In the embodiment shown, a milking station is shown very schematically in which a cow 2 is milked by a milking robot 6 using teat cups 7 (one of which is shown here). Once the milking process starts, i.e. after the cow has been identified and it has been decided that she is to be milked, the control device 4 will be able to roughly estimate the time that the cow will spend at the milking station. For this purpose, it can use a standard minimum time, the historical milking times of the cow, or even an estimated milking time based on production and milking intervals, as is known per se in the art. For almost every milking, this time is at least 5 minutes, and usually up to 8 or 9 minutes. In the event that the cow is not milked, she will be driven outside and there will not be enough time for a meaningful measurement.

[0046] Alternatively, the station is a feeding station, schematically indicated by a feed trough 8, which has a sensor 9, which indicates the start of feeding by the pressing of the cow's nose. Most milking stations will also have a feed trough, but the feed trough may also be a separate system. Further alternatives may be drinking stations (or drinking troughs), treatment stations, etc. Moreover, most stations have an animal identification (not shown here), by which the animal can be identified and the settings (milking, feeding, treatment) can be adjusted individually. Furthermore, the 3D camera system can also be triggered by such an animal identification, for example in case the system should only monitor specific animals.

[0047] When the 3D camera is turned on, it starts imaging the region of interest (hereinafter: ROI) of the animal, where the region of interest is indicated as the left lumbar fossa (indicated by the dotted line). In order to be able to obtain sufficiently reliable data, the frame rate is at least one frame per second, but preferably at least ten times that. In order to ensure that the ROI is within the field of view of the camera during imaging, a wide-angle lens can be provided on the camera so that when the animal is at or in a (milking, feeding, ...) station, the ROI will be within the field of view. A motor for moving the camera based on the recognition of the ROI in the image by the image processing device 5 can also be provided. Tracking the ROI in this way is a known technique per se. The advantage is that the ROI can form a larger part of the image and can therefore be imaged with a higher resolution.

[0048] The obtained 3D image is processed by the image processing device 5, which will be explained further below. The result of this processing is the curvature value of the ROI as a function of time. This function is analyzed by the control device 5 and one or more criteria are applied to determine the health indication of the cow 2. In the case of health indications causing alarms, etc., the control device 4 can add the cow 2 to a watch list, issue an audible or visual alarm, isolate the cow 2 after the station 2, 6, etc. The cow 2 will then be further examined by the farmer or veterinarian.

[0049] When using the system and in the method, the 3D image obtained forms a 3D representation of the ROI. In order to limit the number of calculations, as will be described below, the ROI can be restricted to only a portion of the left paralumbar fossa, such as to a sub-region or window 11 in this example, although this is not necessary.

[0050] Instead of determining the depth or volume of the ROI (as is done in prior art systems to determine rumen filling), the present invention determines a changing curvature value for the ROI. This is based on the insight that natural processes influence rumen shape in such a way that the predictive value of instantaneous rumen filling values ​​appears to be limited, but instead the predictive value of an analysis of the temporal changes in the curvature of the rumen or left paralumbar fossa region appears to be meaningful.

[0051] For each acquired 3D image, the region of interest (sub) is tracked by image processing (e.g., by identifying the upper left corner of the left paralumbar fossa) and repositioning / adjusting the image size. Then, the curvature value is calculated. This can be done in a variety of ways, as long as it expresses the curvature of the ROI in a systematic way. An example will now be briefly described.

[0052] The determination of the curvature will be limited to the sub-region 11. For this sub-region 11, the curvature value is determined as follows.

[0053] First, the covariance matrix is ​​calculated from the nearest neighbors of the point.

[0054]

[0055] Where k is the number of neighboring points, p i is the position vector of the ith neighboring point, and is the position vector of the centroid of the neighboring points. The resulting covariance matrix C will be a 3×3 matrix with 3 eigenvalues. The surface curvature σ can be estimated by the following equation

[0056]

[0057] Among them, λ0, λ1, λ2 are the eigenvalues ​​of the covariance matrix C, among which λ0 is the smallest eigenvalue.

[0058] The resulting curvature value σ can then be plotted as a function of time. Figure 2, as a somewhat variable curve. Note that the x-axis represents the image number and that in this case the frame rate is 30 Hz, or 1800 images / minute. Obviously, while a rough "beat" can be discerned in the curve, it is difficult to extract meaningful information from it. However, it has been realized that various sources of noise can be effectively eliminated. On the one hand, animal movement can be removed, as well as changing lighting conditions, which in principle are one-off changes rather than regular changes. In addition, regular movements that are much faster than the expected contraction frequency can be eliminated, such as breathing. The latter are typically between about 25 and 50 breaths per minute, which is an order of magnitude higher than rumen reticulum contractions. In this case, the above-mentioned "noise signal" is removed by decomposing the signal by its frequency, where the "pass" frequency range is between 0.5 and 4.0 contractions per minute, while the rest of the frequencies are discarded. To do this, a Fourier transform of the signal is constructed, this pass frequency range and a cutoff frequency range below and above this range are applied, and an inverse Fourier transform is constructed to regenerate the curvature-time function. This function is also plotted in Figure 2 , as a smoothed curve. In the smoothed curve, the more or less evenly spaced peaks are clearly visible, and the frequencies of these peaks are clearly within the expected range. Note that the smoothed curvature function is only used for visual inspection, while the Fourier transformed function is used as the basis for calculations and monitoring.

[0059] In the smoothed curve, the peaks or maxima have been indicated with dots. Thus, for rumen-reticulum contractions, the average frequency is about 9 / (5800 / 1800) = 2.8 contractions per minute. This is a normal frequency for the cows being tested, in this case healthy primiparous Holstein and Brown Swiss cows, so for this particular cow, a "healthy" indication can be given and no health alert needs to be given.

[0060] However, it is possible that, based on historical measurements, a particular cow usually has higher or lower values. In this case, a health indication of "still healthy, but please check" can be given, i.e., some examination can be carried out, but it is not urgent, or the cow can be monitored more closely. It is also possible that the calculated contraction frequency is actually lower than the predetermined threshold, such as when the frequency determined during feeding is between about 1 and 2 per minute. Such values ​​are not uncommon during resting periods, but should usually be higher during feeding. Therefore, in this case, a health indication of "urgent check" can be issued by the control device, or an alarm can be sounded, etc.

[0061] In the example shown, a 3D camera system is provided in a milking station. In this station, the animals are milked several times a day, such as 2-4 times / day. It is advantageous if the animals are monitored more frequently, because, for example, measurements may fail due to being too short to yield meaningful information, or due to too much movement, etc. Therefore, it is advantageous if the 3D camera system includes one or more additional cameras in other locations such as feeding stations, watering stations, or even stalls or "resting stations". Note that the determined frequency can be compared with a correspondingly adjusted threshold frequency (such as a lower threshold frequency when in a stall).

[0062] Note that a 3D camera generates 3D images that represent 2D images of an animal in which each pixel is combined with information about the distance to the camera. In the above-described embodiment, the curvature is calculated in the region of interest, and a conclusion is drawn based on a temporal analysis of the curvature. It is also possible that a control device having an image processing device calculates the distance between a specific point in the region of interest and the camera. The point also moves toward the camera and back again at the same frequency as the curvature changes. In other words, the relative position of the point relative to the camera changes at the same relevant frequency. Therefore, the present invention - both the system and the method also work when they (are arranged to) measure the relative position of a fixed point in the region of interest (the left lumbar fossa or a sub-region thereof), and analyze the extreme values, in particular the time point at which the fixed point is closest to the camera.

[0063] Advantageously, fixed points are determined with sufficient precision and accuracy to prevent human factors or simple erroneous measurements. In addition, advantageously, fixed points can be easily identified in images. In this article, it is helpful to determine the border of the lumbar fossa, which is a relatively easily identifiable triangle on the left side of the animal. Then, the relative position of the border of the region of interest determined in this way can be determined by image processing software, such as in the geometric center or any other position. Then, it is relatively simple to determine the basic position of the animal by using the relative position of the border of the region, even after the fact rather than when analyzing in real time (such as using ribs and spine, which additionally shows that determining the respiratory rate is advantageous, so as to subtract corresponding signals). The movement of these borders as a whole is counted as the displacement of the animal as a whole, and is meaningless for contraction. After subtracting these or otherwise affecting the signal, the real relative motion of the fixed point can be determined, and the remaining analysis can be more or less replicated for this relative motion.

[0064] The above embodiments are only used to help illustrate the present invention and are not intended to limit the present invention in any way. The scope of the present invention is determined by the appended claims.

Claims

1. A system for automatically monitoring ruminants, the system comprising: - a 3D camera system for obtaining a plurality of 3D images of at least one region of interest of said animal at successive points in time during at least a predetermined period of time, - a control device connected to the 3D camera system, the control device being provided with - an image processing device for processing the obtained plurality of 3D images, and - output devices, The control device is arranged to determine rumen motility and health indications based on the processed images and output the health indications to the output device, Wherein, the image processing device is arranged as - determining a region of interest in the plurality of 3D images, and One of the following: - calculating a curvature value of the region of interest for each 3D image and determining the calculated curvature as a function of time, or - measuring for each 3D image the relative position of at least one predetermined point in the region of interest with respect to the animal and determining the relative position as a function of time, Wherein the control device is arranged to determine: - a plurality of time points at which a plurality of local extrema of the function of time occur, and - said rumen motility and said health indicator, wherein determining said rumen motility and said health indicator is performed by analyzing said plurality of time points relative to a predetermined standard.

2. The system according to claim 1, wherein: The ruminant is a dairy animal.

3. The system according to claim 2, wherein: The dairy animals include cows.

4. The system according to claim 1, wherein: The 3D camera system includes a time-of-flight camera or a structured light camera.

5. A system according to any preceding claim, wherein: The region of interest is the left paralumbar fossa.

6. The system according to claim 5, wherein: The region of interest is a partial area within the left paralumbar fossa.

7. The system according to any one of claims 1 to 4, wherein: The curvature value includes an average value of curvature values ​​of the region of interest.

8. The system according to claim 7, wherein: The average value of the curvature value of the region of interest is an average value obtained by averaging over a plurality of points across the region of interest.

9. The system according to any one of claims 1 to 4, wherein: The control device is arranged to determine a frequency of the plurality of local extrema over time and to analyse the plurality of time points by comparing the determined frequency with a predetermined frequency criterion.

10. The system according to claim 9, wherein: The control device is arranged to filter the function of time of the calculated curvature by filtering out time variations of the function having frequencies outside a predetermined frequency range.

11. The system according to claim 10, wherein: The control device is arranged to: - performing a Fourier transform on said calculated curvature function to construct a transformed frequency function, - removing all parts of the transformed frequency function outside the predetermined frequency range to obtain a clean frequency function, and - Determine the contraction frequency by analyzing the clean frequency function.

12. The system according to claim 11, wherein: Determining the contraction frequency by analyzing the clean frequency function includes determining the contraction frequency by determining a frequency value having a strongest signal value or an average value of the frequency signal within the clean frequency function.

13. The system according to claim 10, wherein: The predetermined frequency range is or includes frequencies from 0.5 to 4 times per minute, inclusive.

14. The system according to claim 9, wherein: The frequency criteria includes generating a health warning when the determined frequency is below a predetermined frequency threshold.

15. The system of claim 14, wherein: The frequency threshold is a historical value for the animal, or wherein the frequency threshold is a literature value for the animal.

16. The system of claim 15, wherein: The frequency threshold is a historical average for the animal.

17. The system of claim 15, wherein: The literature values ​​for the animal depend on one or more of the breed, age, days in lactation, or feed type or feeding regimen of the animal, respectively.

18. The system of claim 14, wherein: The frequency threshold depends on the type of activity performed by the animal.

19. The system of claim 18, wherein: Types of activity include feeding, milking, or resting.

20. The system according to any one of claims 1 to 4, comprising at least one station, each station being selected from the group consisting of: a milking station, a feeding station, a drinking station, a treatment station, an isolation station and a selection station, in, The 3D camera system includes a 3D camera disposed in at least one of the stations.

21. The system of claim 1, wherein: The control device is arranged to determine the time intervals between the plurality of local extrema based on the determined plurality of time points and to analyze the plurality of time points by comparing the determined time intervals with a predetermined time interval criterion.

22. A method for determining an indication of health of a ruminant, the method using a system according to any one of the preceding claims and comprising the steps of: - obtaining a plurality of 3D images of at least one region of interest of said animal at consecutive time points during at least a predetermined time period, - Processing of the multiple 3D images acquired, - determining a region of interest in the plurality of 3D images, One of the following: - calculating a curvature value of the region of interest for each 3D image and determining the calculated curvature as a function of time, or - measuring for each 3D image the relative position of at least a predetermined point in the region of interest with respect to the animal and determining the relative position as a function of time, And determine: - a plurality of time points at which a plurality of local extrema of the function of time occur, and - Rumen motility and said health indicator, wherein determining said rumen motility and said health indicator is performed by analyzing said plurality of time points relative to a predetermined standard.

23. The method of claim 22, wherein the ruminant comprises a dairy cow.

24. A method for monitoring the health of a ruminant, the method comprising: - automatic determination of the rumen-reticulum contraction frequency and rumen filling by a system according to any one of claims 1 to 21, and A health warning is generated based on a comparison of the determined rumen-reticulum contraction frequency and the determined rumen filling with a predetermined frequency value and a predetermined rumen filling value, respectively.

Images